As computational resources continue to evolve, NETL researchers look to new and more powerful tools to bolster their ability to model complex fossil energy power systems. The Lab has decades of experience developing this kind of software — known as computational fluid dynamics (CFD) code — including the award-winning Multiphase Flow with Interphase eXchanges (MFiX). In recent years, machine learning (ML) techniques have been integrated into powerful computational platforms like Google’s Tensor Flow, which is revolutionizing the way NETL researchers write CFD code to accelerate the design of more energy-efficient systems.

In 1975, Gordon Moore, an engineer and businessman, made an important prediction about the future of computational development. Moore’s law, as it is known today, observes that the number of transistors on integrated circuit chips doubles about every two years, which in turn guides nearly all technology development. Still today, computational power is not only increasing, it’s also diversifying, and researchers are using a variety of processing tools including computer processing units (CPUs) and graphic processing units (GPUs), the latter of which allow more calculations to be made in parallel.

“That is where TensorFlow comes into play,” NETL’s Dirk VanEssendelft, Ph.D. said. “It is a machine learning framework that takes care of memory management, communication, data operations, optimization and parallelization, so we can just focus on the algorithm. There’s no more worrying about rewriting code for specific hardware. It’s basically a practical and easy on-ramp for our code to run on the best hardware available.”

VanEssendelft works on NETL’s multiphase flow science team, which investigates the flow of different phases of matter in energy systems using powerful CFD software like MFiX. He and his team have already had great success with TensorFlow.

“TensorFlow allowed us to leverage machine learning approaches to overcome one of the biggest drains to computational resources when modeling advanced energy systems,” VanEssendelft said. “Our Machine Learning Accelerated Stabilized Explicit Variable Load software uses Tensor Flow to solve complex chemical reaction equations much faster than previous iterations, drastically shortening design time and significantly reducing research and development costs.”

The NETL team recently developed a single-phase CFD code in TensorFlow, running on GPUs, in just a few months. This code is helping to solve the fluid portion of their modeling efforts. Additionally, the researchers are using TensorFlow to work on incorporating a ML training engine to run alongside a CFD solver engine, helping to further refine predictions of the CFD code.

Just as Moore foresaw the inevitable explosion of computational power, NETL recognizes the ever-increasing possibilities that AI and ML are affording researchers. Researchers like VanEssendelft and his team are unlocking the potential of ML frameworks to accelerate their own research, which will bring the nation more affordable and reliable energy.